server : merge split UTF-8 token text in verbose JSON

[lyonsno]

Fixes #3821.

This applies the UTF-8 token-boundary merge used by the CLI JSON path to the server verbose_json word output. The UTF-8 tail detector now lives in examples/common-whisper, so both CLI and server use the same helper.

When a multi-byte character is split across adjacent tokens, the server now emits one merged segments[].words[] entry instead of separate invalid UTF-8 fragments. segments[].tokens still preserves the raw non-EOT token ids, so a merged word can correspond to more than one token id.

For merged words, metadata follows the existing CLI lead-token convention: start, probability, and t_dtw come from the lead token, while end advances to the last consumed token with a valid end time.

Verification:

• cmake --build build-3821 --target test-common-utf8 whisper-server whisper-cli test-vad -j 4
• ctest --test-dir build-3821 -L unit --output-on-failure
• ctest --test-dir build-3821 -R '^test-whisper-cli-tiny\.en$' --output-on-failure
• git diff --check